[Core] Improve Tensor serialisation

[lgeiger]

This improves v1 tensor serialisation by directly relying on torch.frombuffer which removes the need for an temporary numpy array which is a bit easier to read.

Here's a small micro benchmark to verify that this is also faster:

import numpy as np
import torch

from vllm.v1.serial_utils import MsgpackDecoder, MsgpackEncoder

encoder = MsgpackEncoder()
tensor_decoder = MsgpackDecoder(torch.Tensor)
numpy_decoder = MsgpackDecoder(np.ndarray)

array = np.random.rand(1, 3, 896, 896).astype(np.float32)
tensor = torch.tensor(array)
encoded_tensor = encoder.encode(tensor)
encoded_array = encoder.encode(array)

main:

Tensor encode/decode
5.44 μs ± 12.3 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
5.22 μs ± 9.13 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
Array encode/decode
5.46 μs ± 4.99 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
1.43 μs ± 4.35 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

This PR:

Tensor encode/decode
5.03 μs ± 3.81 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
4.41 μs ± 7.52 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
Array encode/decode
5.02 μs ± 23.8 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
1.45 μs ± 3.2 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[njhill]

Thanks @lgeiger, LGTM

[njhill]

I'm going to revert this since it broke some tests: #18857. We can look into the reason and open another PR as needed.

[lgeiger]

Oh so sorry, about that. I can have a look at it in a bit

[lgeiger]

#18860 Should fix it